Problem Solver

How to Build a Human-Level Intelligence

Step 1: A Goal-Directed Analogy-Based Problem Solver

April 27, 1998

This is from a research summary I sent to the SARA98 symposium on reformulation.

My goal is to build a cognitive architecture based on Marvin Minsky’s "Society of Mind" theory [1]. I think of a cognitive architecture as a system that consolidates the cognitive functionality required for ordinary commonsense problem solving. Here are some of the functions that must be served:

Recognizing Abstracting Comparing Expecting Learning Predicting Explaining Planning

There have been no successful attempts to integrate all of these capabilities into a single AI system (or even to settle on what they are! A longer list is available here, though it is far from complete.) We must do this for artificial intelligence to move to the next level: human-level versatility, resourcefulness, and robustness.

How can we do this so that the resulting integration is tight and robust? Present AI systems can be thought of as being made of the simplest agents—usually there is a fixed algorithm for matching, planning, credit assignment, etc. But if those subfunctions were themselves served by simpler AI systems, then the overall system could be much more flexible, resourceful, self-improving, and internally well-coordinated. I want to explore how to build a new kind of AI system, one where the subfunctions themselves are managed by goal-directed problem solvers.

To this end I am building a fast goal-directed analogy-based problem solving system that will serve as the base level. Its main feature is that representations are multiple, corresponding, and parallel.

Multiple Representations

The system is being designed from the beginning to exploit many representations. I believe that assuming a fixed representation of the situation is at the root of the failure of many AI techniques to scale up to more realistic domains. For example:

	Learning algorithms are always presented with the right features, rendering their task obvious.
	Planning systems are given complete models of their operators, and so are never surprised by their actions having unexpected effects.

Reasoning systems assume that the world contains only a well-defined set of objects, and so do not have to take into account contingencies generated by sources they have not considered.

And so forth. Many researchers now avoid using explicit representations and a movement has arisen to try to build systems with as little representation as possible, for example the robots in Rodney Brooks’s lab [2].

I believe that we should do just the opposite. The simplicity of representations is not a weakness but a strength. Ideally they make explicit only the important aspects of the situation. An AI system can be made robust by providing it with many representations and the capacity to reformulate between them, as work on the problem progresses and as what is considered to be important changes. If a representation begins to fail we can switch to another better one. There is no need for the system to have a perfect model of the problem situation if is provided with enough imperfect models and the ability to combine or move between them.

Corresponding Representations

AI systems should be able to learn from all of their experiences, so I am building into the system the ability to exploit knowledge existing anywhere in the system. Much of our conceptual universe is constructed by finding correspondences between representations of different varieties. George Lakoff and Mark Johnson have convincingly argued that the knowledge and skills we have for reasoning about space and time are also used to help reason about social realms [3]. This general idea, when applied between arbitrary different representations in the system, allows for extremely fast learning and great resourcefulness, for then any one agent in an AI system can draw not only from its own past experiences, but from the knowledge and experiences of other agents.

Building correspondences between representations requires finding the analogies and isomorphisms between them. I am designing the system so that it is easy to organize representations into societies where each representation is capable of describing the same thing in different ways for similar purposes.

Parallel Representations

A third feature of the system is the ability to exploit parallelism. There are many potential benefits from moving beyond merely switching between representations to using many representations at the same time:

	Speed. Problems can be solved faster with N representations than with one—but how much faster? We suspect that in some cases we can actually do much than a factor of N, because if we try not to deactivate representations we do not lose state, and so do not have to reconstruct lines of reasoning or have to repeat negative experiences.
	Eases reformulation. Having many representations active simultaneously makes it easier to invoke new ones! To use a new representation we must first be able to compute some of the values of its descriptors. The more representations are active, the more paths and intermediate results are available for computing those values, and so for using new representations.
	Robustness. Any particular representation is bound to be broken in some respect, or incomplete in its knowledge. Bringing many different representations to bear on the same problem at the same time lets us simultaneously apply a diverse variety of constraints towards the solution.

Summary

The kinds of problem solving systems produced so far in AI are not the final product, but only ways to build agents of even bigger systems. Designing "functional sketches" of cognitive architectures and finding ways to build them by piecing together smaller goal-directed problem solvers is the next big challenge in artificial intelligence. But to do so we must first build robust problem solving systems, which will require special facilities for using multiple representations.

References

[1] Marvin Minsky. The Society of Mind. Simon and Schuster: New York, New York. 1986.

[2] Rodney Brooks. Intelligence Without Representation. Artificial Intelligence, 47, 1991, 139—160.

[3] George Lakoff and Mark Johnson. Metaphors We Live By. University of Chicago Press: Chicago, Illinois. 1980.

Bibliography

Push Singh. Failure-Directed Reformulation. M.Eng. Thesis. MIT Department of Electrical Engineering and Computer Science. 1998.